Kernel card: padded-fp8-producers

Purpose

Remove separate normalization, activation, quantization, and row-padding launches before fixed-tile FP8 GEMMs.

Numerical contract

• BF16 normalization and modulation seam before FP8 conversion.
• E4M3FN saturation range [-448, 448].
• FP16 merged SwiGLU input still rounds the activation product through BF16 before static FP8 quantization, matching the production model seam.
• Padded rows are exactly zero.
• Residual input is not mutated; the updated BF16 residual is returned through residual_out.

Functions

• adaptive_rms_norm_quant_fp8_padded_bf16
• residual_add_adaptive_rms_norm_quant_fp8_padded_bf16
• swiglu_quant_fp8_padded_bf16
• swiglu_merged_quant_fp8_padded_bf16
• swiglu_merged_quant_fp8_padded_fp16

Scope

CUDA 12.8 or newer. This package contains CUDA kernels and does not claim ROCm support. Shape and architecture claims are limited to tested build artifacts.